After ~1.5 weeks uptime healthchecks of local web-server start showing errors “Too many sockets on the process”
It turned out that haproxy process uses 20000+ more file descriptors that it usually needs (in compare to another similar servers with 1.7.9 and much longer uptime).
According to /proc/net/sockstat it holded 10000+ “inuse” UDP-sockets.
Outgoing UDP-connections is used in LUA, but very rare.
Http/2 is disabled, reloads are used very often
Haproxy restart solved the problem and looks like now amount of UDP-sockets is slowly growing again